AgentsMesh: the repo is the harness — and three parallel streams is a hard ceiling
Mar 4 AgentsMesh field note: 52 days / 600 commits building the fleet console with itself; Loops (cron) vs Tickets; tech-debt amplification; Autopilot after cognitive bandwidth broke at ~3 workstreams.
Source: Building AgentsMesh with AgentsMesh: 52 Days of Harness Engineering by One Person
AgentsMesh’s Mar 4 field note (HN follow-up Mar 14) is a self-bootstrap receipt: one person, 52 days, 600 commits, ~966k lines of throughput and ~356k still standing — building a multi-machine agent fleet console with that console. The argument is not the line count (git log is public); it is that the engineering soil sets the agent ceiling. Layered DDD modules, directory names as navigation maps, strong typing as a compile-time gate, and a four-layer feedback loop (hot reload → unit → e2e → CI) matter more than prompt craft because agents amplify whatever patterns dominate the tree — including tech debt. Temporary shortcuts stop being local; they become the precedent the next agent copies.
Two operational claims land for continuous factories. First, decomposition shapes: Tickets for one-shot work with kanban/MR tracking; Loops for recurring cron work (daily tests, scheduled builds, quality scans) with independent LoopRun records — the difference between a burst session and an always-on factory schedule. Second, cognitive bandwidth is a hard constraint: around day five, three parallel worktrees was the sound-decision ceiling; a fourth visibly degraded architecture. The escape was Autopilot — a control agent that keeps a pod moving with iteration caps and human takeover — so the human supervises the supervisor system, not every terminal. Isolation (Pods/worktrees), Channels, and Bindings (pod:read/pod:write) are the coordination primitives once you leave single-agent heroics.
Why it matters for code review
Steady harnessed throughput compounds only when recurring verification is a first-class Loop, not a morning ritual you remember, and when parallel agent load stays inside a judgment budget a human can actually hold. Burst Claude mastery still wins on greenfield taste; under sustained fleet load the review artifacts are the LoopRun records, the amplified-debt cleanups, and whether Autopilot escalations are legible — not how fast a single session typed.